Update app.py

app.py

@@ -23,7 +23,6 @@ CSS = """
 
 # Ensure model and scheduler are initialized in GPU-enabled function
 unet = UNet2DConditionModel.from_config(base, subfolder="unet").to("cuda", torch.float16)
-unet.load_state_dict(torch.load(hf_hub_download(repo, checkpoints)), map_location="cuda")
 if torch.cuda.is_available():
     pipe = DiffusionPipeline.from_pretrained(base, unet=unet, torch_dtype=torch.float16, variant="fp16").to("cuda")
 
@@ -38,6 +37,7 @@ def generate_image(prompt, ckpt):
     num_inference_steps = checkpoints[ckpt][1]
 
     if loaded != num_inference_steps:
+        unet.load_state_dict(torch.load(hf_hub_download(repo, checkpoints)), map_location="cuda")
         pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample" if num_inference_steps==1 else "epsilon")
         loaded = num_inference_steps